bzdww

Get answers and suggestions for various questions from here

[Technical Review] Research on Facial Expression Recognition

cms

The article was first published on the WeChat public account "There are three AI"

[Survey] facial expression recognition research mp.weixin.qq.com

With the rapid development of machine learning and deep neural networks and the popularity of smart devices, face recognition technology is experiencing unprecedented development, and the discussion on face recognition technology has never stopped. At present, the accuracy of face recognition has surpassed that of the human eye. At the same time, the basic conditions of software and hardware for mass popularization are already available, and the application market and field demand are very large. The market development and specific application based on this technology are showing a booming trend. Facial expression recognition (FER) is an important part of face recognition technology. In recent years, it has gained extensive attention in human-computer interaction, security, robot manufacturing, automation, medical, communication and driving. Research hotspots in the industry and industry. This article will give a more detailed overview of the relevant content of facial expression recognition in face recognition.

01 expression related overview

1.1 Expression definition and classification

“Expression” is a word that is mentioned in many of our daily lives. In interpersonal communication, people can enhance communication by controlling their facial expressions. Facial expression is an important way to spread human emotion information and coordinate interpersonal relationships. According to the research of psychologist A. Mehrabia, in human daily communication, the information transmitted through language only accounts for 7% of the total amount of information. The information transmitted by facial expressions reaches 55% of the total amount of information. It can be said that we display our own expressions every day and receive other people's expressions. What is the expression?

Facial expressions are the result of one or more actions or states of facial muscles. These movements express the emotional state of the individual to the observer. Facial expressions are a form of nonverbal communication. It is the primary means of expressing social information between humans, but it also occurs in most other mammals and other animal species.

There are at least 21 kinds of human facial expressions, in addition to the common six kinds of happiness, surprise, sadness, anger, disgust and fear, as well as 15 kinds of compound expressions that can be distinguished, such as surprise (happy + surprised), grief (sad + anger). .

The expression is the emotional index projected by humans and other animals from the appearance of the body. Most of them refer to the state of facial muscles and facial features, such as smiles and anger. It also includes the body language that the body as a whole expresses. Some expressions can be accurately explained, and even between members of different species, anger and extreme satisfaction are the main examples. However, some expressions are difficult to explain, and even between familiar individuals, disgust and fear are the main examples. In general, the various organs of the face are an organic whole, and the same emotion is expressed in a coordinated manner. Facial expressions are part of the body's (physical) language and are a physiological and psychological response that is often used to convey emotions.

1.2 Research on expressions

The study of facial expressions began in the 19th century. In 1872, Darwin elaborated on human facial expressions and facial expressions of animals in his famous book The Expression of the Emotions in Animals and Man (1872). The link and difference between.

In 1971, Ekman and Friesen pioneered the work of modern facial expression recognition. They studied the six basic expressions of human beings (ie, happiness, sadness, surprise, fear, anger, disgust), determined the categories of identifying objects, and A database of facial expression images with thousands of different expressions has been systematically established, and the facial changes corresponding to each expression, including eyebrows, eyes, eyelids, lips, etc., are described in detail.

In 1978, Suwa et al. conducted an initial attempt to recognize facial expressions in a face video animation, and proposed automatic facial expression analysis in the image sequence.

Beginning in the 1990s, K.Mase and A.Pentland used optical flow to determine the main direction of muscle movement. After using the proposed optical flow method for facial expression recognition, automatic facial expression recognition entered a new era.

1.3 micro-expression

With the deepening of the study of expressions, scholars have focused their attention on the study of a more subtle expression, that is, the study of micro-expressions. So what is a micro-expression?

A micro-expression is a psychological noun. It is a short-lived facial expression that an unconscious person makes when trying to hide an emotion. They correspond to seven world-wide emotions: disgust, anger, fear, sadness, happiness, surprise, and contempt. The duration of the micro-expression is only 1/25 second to 1/5 second, expressing the true emotion that one tries to suppress and hide. Although a subconscious expression may only last for a moment, sometimes it expresses the opposite emotion.

Microexpressions have great commercial and social significance.

In the United States, research on micro-expression has been applied to areas such as national security, the judicial system, medical clinical, and political elections. In the field of national security, some dangerous people such as trained terrorists may easily pass the detection of polygraphs, but through micro-expressions, they can generally find their true expression under the false surface, and because of the characteristics of micro-expressions It also has a good application in the judicial system and medical clinic. Film producer directors or advertisement producers can also predict how the proceeds of the promo or advertisements will be based on the crowd sample collection method to watch the video or the micro-expressions of the advertisements.

In short, with the advancement of science and technology and the continuous development of psychology, the research on facial expression will be more and more in-depth, the content will be more and more rich, and the application will be more and more extensive.

02 Expression recognition application

2.1 Online API (1) Microsoft Azure
This API includes face verification, face detection, and expression recognition. For the face recognition function integrated with the face API, the confidence level can be returned for a series of expressions (such as anger, contempt, disgust, fear, happiness, no emotion, sadness, and surprise) on all faces on the image, and the recognition result is returned through JSON. These emotions can be thought of across cultural boundaries and are usually conveyed by specific facial expressions. Links: azure.microsoft.com/zh- Figure 2.1 shows the results of the face API recognition.

Figure 2.1 Microsoft Azure face API expression recognition actual operation diagram (2) Baidu AI open platform (with WeChat applet)
This API can detect the face in the picture and mark the face for the face. After detecting the face, the face can be analyzed, and 72 key points such as eye, mouth and nose contours can be obtained to accurately identify various face attributes, such as gender, age, expression and the like. This technology can adapt to a variety of real-world environments such as large-angle side faces, occlusion, blur, and expression changes. Link: ai.baidu.com/tech/face/ Figure 2.2 shows a functional demonstration of the API.

Figure 2.2 Functional demonstration of the Baidu AI Open Platform Face API

(3) Tencent excellent map AI open platform (with WeChat applet) The API uses a smart strategy to search for any given image to determine whether it contains a human face, and if so, returns the location of the face, Size and attribute analysis results. Currently supported face attributes are: gender, expression (neutral, smile, laughter), age (error estimate is less than 5 years old), whether to wear glasses (ordinary glasses, sunglasses), whether to wear a hat, whether to wear a mask. At present, the excellent face detection and analysis is not only mature applied to the face value analysis in the picture, but also to start the camera and other entertainment scenes when the face is detected, and can also be easily detected and counted by the face in the image or video. Understand the flow of people in the region, and through the face detection and analysis of the advertising audience, to understand the gender, age and other attributes and distribution of the crowd, according to which to more accurately match the advertising. Link: youtu.qq.com/# Figure 2.3 shows a functional demonstration of the API.

Figure 2.3 Tencent Uto AI Open Platform Face API feature demo

2.2 APP

(1) Polygram
WeChat has become an indispensable part of our life. Social, transfer, payment, and shopping can all be used as carriers. WeChat, brushing friends, sending a message, and fighting pictures have become the daily routine of our free time. Various WeChat emoticons have become a major mainstream. Polygram is different from previous social software. It is a social software based on face recognition emoticons. It adds face recognition and neural network technology. It can use the user's facial expression to generate an emoji. Here, the user can search and send the corresponding expression through face recognition technology. Polygram is an artificial intelligence dynamic social network that understands facial expressions. It is mainly characterized by face recognition based expression pack, which is capable of detecting the real expression of the face by using face recognition technology, thereby searching for the corresponding expression and transmitting the expression. When a user posts a picture or video on a Polygram, it is very smart to use facial recognition technology and a mobile phone camera to automatically capture the realities of the face when the user browses photos, text, videos, etc. shared by friends on the social platform. Expressions, you will understand how your friends feel about them. This is done by mimicking the facial emoji of facial expressions and allows the user to react to their faces.

Figure 2.4 User is using Polygram

(2) The emo
emo is a music app that can recognize emotions. We always find the songs in the playlist after opening the phone to open the music player, but it is difficult to save hundreds of songs. Finding the moment you want to listen to the playlist is not due to obsession, just because of your mood. When you are happy, you want to listen to the song of jumping; when you are sad, you should let the songs of the lower ones; when you are excited, you need to be passionate. When everyone has different moods, everyone needs different music antidote. Emo is born, solves the troubles of listening to songs, and plays the most suitable songs at the most suitable time.

In front of you in emo, you will be the most honest, you don't have to hide your mood, happy is happy, sadness is sadness. Emo will scan your face through the front camera and calculate your current state of mind. You will be amazed at its high accuracy, and it is not only happy and sad, it can also "see" other moods such as: calm, Confused, surprised, angry, and so on.

Estimating mood is not the only amazing place. After deducing your mood, emo will push music for you. Emo has a huge collection of high-quality music background songs. Each song that is pushed is manually labeled with a mood. Each song is carefully selected by you for your mood. Simply put, emo is a music player, and the embedding of face recognition technology makes this player not so simple - emo can scan the user's facial expressions, judge the user's emotions, and recommend the user to the corresponding music. The idea of ​​the product is to hope that the user can hear the songs that match the mood at every moment. Overall, the app also jumped out of the player in the general sense, is a very interesting product, looking forward to better optimization. The other three major music players may be able to learn from the future.

2.3 Analysis summary

At present, the APIs of major manufacturers are very mature. At the same time, due to the rise of WeChat applets, many APP functions can be migrated to small programs. Through extensive research, it can be found that there are many products for face recognition. Focusing on facial expression recognition is not much, or simply giving simple expressions such as whether to smile or not, most of them do not have an organic combination with the product. During the research process, I personally felt that emo is a good idea, but unfortunately it has not been well promoted.

At present, only the technology for face recognition is relatively mature, and there is still a large market for expression recognition. What needs to be done is to apply expression recognition to the actual scene and combine it with the actual needs. For example, in the production of the game, real-time reflection can be made according to human emotions to enhance the player's immersion; in the distance education, the progress of the lectures, teaching methods, etc. can be adjusted according to the expression of the students; in the aspect of safe driving, the driver can be judged according to the expression of the driver. Driving state to avoid accidents. In terms of public safety monitoring, it is possible to judge whether there is abnormal emotion or crime prevention according to the expression; when making commercials, the producer often has a headache: when to insert the trademark logo, when to jump out of the product image can The audience has a deeper impression of this brand and this product? Expression recognition can help ad producers solve this headache. The author only needs to invite some people to try this commercial after the commercial is completed, and use the expression recognition system to test the viewer's emotional changes during the trial, and find the paragraph with the most emotional fluctuations. This is the best logo insertion. paragraph. Similarly, it can help ad producers find the best logo placement point, and also help filmmakers find the most attractive part of a movie to make a movie trailer to ensure that the trailer is attractive enough. To ensure that more people are willing to walk into the cinema to watch the "positive film" after watching the trailer. Expression recognition is a promising direction. Linking it to everyday needs is an important factor to consider for such products, not just a test result, perhaps one of the future directions.

03 expression commonly used open source database

(1) KDEF and AKDEF (karolinska directed emotional faces) data sets

Link: emotionlab.se/kdef/ This data set was originally developed for psychological and medical research purposes. It is mainly used for experiments such as perception, attention, emotion, and memory. In the process of creating the data set, the use of uniform, soft lighting, the collector is wearing a uniform T-shirt color. This data set consists of 70 people, 35 men, 35 women, aged between 20 and 30. No beards, earrings or glasses, and no obvious makeup. 7 different expressions, each with 5 angles. A total of 4,900 color maps. The size is 562*762 pixels. Figure 3.1 is an example of a smile in the data set.

Figure 3.1 Example of a smile in KDEF and AKDEF Dataset

(2) RaFD data set
link: socsci.ru.nl:8180/RaFD2 The dataset is organized by the Nijmegen Institute for Behavioral Sciences at Radboud University, a high-quality facial database with a total of 67 models: 20 white male adults, 19 white female adults, 4 white boys, 6 White girls, 18 Moroccan male adults. A total of 8040 pictures, including 8 expressions, namely, anger, disgust, fear, happiness, sadness, surprise, contempt and neutrality. Each expression contains 3 different gaze directions and is simultaneously captured from different angles using 5 cameras. Figure 3.2 is an example of 5 directions in the data set, and Figure 3.3 is an example of an expression in the data set.

Figure 3.2 An example of five directions in the RaFD Dataset

Figure 3.3 An example of an expression in the RaFD Dataset

(3) Fer2013 data set

The data set contains a total of 26,190 48*48 grayscale images, and the resolution of the images is relatively low, with a total of 6 expressions. 0 anger angry, 1 disgust disgust, 2 fear fear, 3happy happiness, 4 sad sad, 5 surprised, 6 normal neutral. Figure 3.4 shows some of the data for the Fer2013 data set.

Figure 3.4 Part of the data for the Fer2013 Database

(4) CelebFaces Attributes Dataset (CelebA) data set

Link: mmlab.ie.cuhk.edu.hk/pr It is a data set used by Shangtang Technology to study face attributes, a large face attribute data set containing more than 200K celebrity images, and each data set has 40 attribute comments. The images in this dataset cover large pose changes and complex backgrounds. CelebA's diversity is very good, with about 100,000 data with a smile attribute, and Figure 3.5 is an example of some smiles in the data set.

Figure 3.5 Some examples of smiles in the CelebA Dataset

(5) Surveillance Cameras Face Database
(SCface)

Links: scface.org/SCface It is a database of static images of faces. The images were taken using five different quality video surveillance cameras in an uncontrolled indoor environment. The database contains 4160 still images of 130 themes (in the visible and infrared spectra). Figure 3.6 is an example of different poses in this data set.

Figure 3.6 Some examples of different poses in SCface Database

(6) Japanese Female Facial Expression (JAFFE) Database

The database contains 213 images of 7 facial expressions (6 basic facial expressions + 1 neutral) consisting of 10 Japanese female models. Each image was rated as 6 emotional adjectives by 60 Japanese subjects. Figure 3.7 shows some of the data in this data set.

Figure 3.7 Part of the data in JAFFE In addition to the open source dataset described above, there are a lot of open source datasets on expressions. In short, we need to search more and summarize. Using these open source datasets, we can save a lot of time for constructing data. It is convenient for us to train a model with better robustness.

04 Facial expression recognition research method

4.1 Expression Recognition System

The facial expression recognition system is shown in Figure 4.1. It is mainly composed of face image acquisition, face detection, feature extraction and feature classification.

Figure 4.1 Human face expression recognition system Because the open source expression database is currently more, the image acquisition is not difficult, the face detection algorithm is mature, and has developed into an independent research direction, so the research of facial expression recognition is mainly reflected in the system. The next two steps: feature extraction and feature classification, the following two steps will be explained from the traditional research methods and deep learning research methods.

4.2 Traditional research methods

4.2.1 Feature extraction

The expression feature extraction mainly adopts mathematical methods, and relies on computer technology to organize and process data of facial expression digital images, extract expression features, and remove non-expression noise. In some cases, the feature extraction algorithm extracts the main features of the image and objectively reduces the dimensionality of the image. Therefore, these feature extraction algorithms also have the effect of dimensionality reduction.
The generation of facial expressions is a very complicated process. If psychological and environmental factors are not taken into account, what is presented to the observer is the simple muscle movement and the resulting changes in facial shape and texture. The static image presents the expression state of a single image when the expression occurs, and the dynamic image presents the motion of the expression between multiple images. Thus when processed according to the state and to differentiate expression occurs, face feature extraction algorithm roughly divided based feature extraction method of a still image and a feature extraction method based moving image . The feature extraction algorithm based on static image can be divided into global method and local method . The feature extraction algorithm based on dynamic image is divided into optical flow method , model method and geometric method .

Feature extraction method based on static image:

(1) overall method
of facial expressions depend on muscles to reflect. The static image of the facial expression visually shows the changes in the facial shape and texture produced by the movement of the facial muscles when the expression occurs. On the whole, this change causes the obvious deformation of the facial organs, which will affect the global information of the face image. Therefore, the facial expression recognition algorithm that considers the expression features from the overall perspective appears.

Classical algorithms in the holistic approach include principal component analysis (PCA), independent component analysis (ICA), and linear discriminant analysis (LDA). Researchers have done a lot of work on this. The literature [1-3] uses FastICA algorithm to extract expression features. This method not only inherits the feature that ICA algorithm can extract hidden information between pixels, but also can complete it through iteration and quickly. Separation of expression features. In [4], the support vector discriminant analysis (SVDA) algorithm is proposed, which is based on Fisher linear discriminant analysis and support vector machine. It can make the expression data have the largest inter-class separation in the case of small sample data. The decision function required to build the SVM algorithm. Experiments show that the recognition rate of this algorithm is higher than PCA and LDA. Literature [5] relies on two-dimensional discrete cosine transform to map face images through frequency domain space, and combines neural networks to classify expression features.

(2)
There is not only an overall change in the facial expression on the static image of the local method , but also a local change. The information contained in the local deformation of facial muscle texture, wrinkles, etc., helps to accurately determine the attributes of the expression. The classical methods of local methods are Gabor wavelet method and LBP operator method. In [6], a variety of feature extraction algorithms such as Gabor wavelet are used as a means to combine the new classifier to experiment on static images. In [7], the 34 face feature points are manually labeled first, then the Gabor wavelet coefficients of the feature points are represented as the marker map vector, and finally the KCCA coefficient between the marker map vector and the expression semantic vector is calculated, thereby realizing the classification of the expression. . The literature [8] proposed the CBP operator method, which reduces the dimension of the histogram by comparing the neighbor pairs of the ring neighborhood. For the modification of the symbol function, the anti-noise performance of the algorithm is enhanced, and the CBP operator method achieves a higher recognition rate.

Feature extraction method based on dynamic image:

The difference between a moving image and a still image is that the moving image reflects the process of the facial expression. Therefore, the expression features of the moving image are mainly manifested in the continuous deformation of the face and the muscle movement in different areas of the face. At present, feature extraction methods based on dynamic images are mainly divided into optical flow method, model method and geometric method.

(1) Optical flow method

The optical flow method is a method of reflecting the gray scale change of corresponding objects between different frames in a moving image. The early facial expression recognition algorithm mostly uses the optical flow method to extract the expression features of dynamic images. This is mainly because the optical flow method has the advantages of highlighting the deformation of the face and reflecting the trend of the face movement. Therefore, the algorithm is still an important method to study dynamic image expression recognition in the traditional method. Literature [9] firstly uses the optical flow field and gradient field between successive frames to represent the temporal and spatial changes of the image, and realizes the expression area tracking of each frame of the face image. Then, the change of the movement direction of the feature area indicates the face muscle. Exercise, which in turn corresponds to different expressions.
(2) Model Method
model facial expression recognition method is the statistical method of expression refers to the video information will be described parameterization. Commonly used algorithms mainly include active shape model method (ASM) and active appearance model method (AAM). Both algorithms can be divided into two parts: shape model and subjective model. In terms of the apparent model, ASM reflects the local texture information of the image, while AAM reflects the global texture information of the image. In [10], a three-dimensional face feature tracking method based on ASM is proposed. This method tracks and models 81 feature points of the face and realizes the recognition of some composite action units. Literature [11] uses the terrain feature model of images to identify facial movements and expressions; uses AAM and manual markers to track facial feature points, and obtains facial expression regions according to feature points; by calculating the terrain histogram of facial expression regions Figure to obtain terrain features to achieve expression recognition. In [12], an AAM algorithm based on two-dimensional apparent features and three-dimensional shape features is proposed. The expression feature is extracted in the environment where the face position is shifted.

(3) Geometry method
In the expression feature extraction method, the researchers consider that the generation and expression of expression are largely reflected by changes in facial organs. The main organs of the face and its pleated parts become areas where the expression features are concentrated. Therefore, marking the feature points in the facial organ area, calculating the distance between the feature points and the curvature of the curve of the feature points becomes a method of extracting facial expressions by geometric form. In [13], the deformation mesh is used to mesh the faces of different expressions, and the coordinates of the mesh nodes between the first frame and the maximum frame of the sequence are used as geometric features to realize the recognition of the expression.

4.2.2 Feature Classification

The purpose of feature classification is to determine the type of expression corresponding to the feature. In facial expression recognition, the category of an expression is divided into two parts: a basic expression and an action unit. The former is generally applicable to all processing objects, the latter is mainly applicable to dynamic images, and the main feature classification methods can be divided into a Bayesian network-based classification method and a distance-based classification method .

(1) Classification method based on Bayesian network

The Bayesian network is a graphical network based on Bayesian formula based on probabilistic reasoning. From the perspective of facial expression recognition, the role of probabilistic reasoning is the process of inferring the probability information of unknown expressions from known expression information. Bayesian network-based methods include various Bayesian network classification algorithms and Hidden Markov Model (HMM) algorithms. In [14], researchers used Naive Bayes (NB) classifier, tree enhancer (TAN) and HMM to implement expression feature classification.

(2) Classification method based on distance metric

The classification method based on the distance metric is to realize the expression classification by calculating the distance between the samples. The representative algorithm has a neighbor method and an SVM algorithm. The nearest neighbor method is to compare the Euclidean distance between the unknown sample x and the samples of all known categories. The distance between the x and the known samples is determined by the distance of the distance. The SVM algorithm finds the different types of samples by optimizing the objective function. The largest hyperplane in the distance between the two. Literature [8] uses the nearest neighbor method to classify expression features, and points out that the shortcoming of the nearest neighbor method is that the size of the classification accuracy depends on the number of samples to be classified. [15,16] proposed improvements to SVM from their respective perspectives. The former combines k-nearest neighbor method with SVM, integrates neighbor information into the construction of SVM, and proposes a local SVM classifier; the CSVMT model proposed by the latter will The combination of SVM and tree modules solves the classification subproblem with low algorithm complexity.

4.3 Deep learning method

The above are some introductions of traditional research methods. The following mainly describes how to apply deep learning to expression recognition. I will take a few articles as an example to introduce the research methods and ideas of the current deep learning method.

Different from traditional method feature extraction, the deep learning method is adopted because the network (especially CNN) in deep learning has better ability to extract features from images, thus avoiding the cumbersome extraction of features. Artificial features include 68 features such as the commonly used Faciallandmarks, and deep learning, in addition to predictions, often plays the role of feature engineering, eliminating the need for manual extraction of features. The following is a description of the types of networks commonly used in deep learning, followed by feature extraction of images through pre-trained networks, and Fine-Tunning for pre-training networks using their own data for fine-tuning.

If the network layer CNN, RNN, Fully-Connect and other layers commonly used in deep learning are combined into a network, a variety of choices will be made. However, the performance of these networks needs to be discussed more and more, after a series of research by many researchers. In practice, many network models already have a lot of performance, such as the model proposed in the ImgeNet competition: AlexNet, GoogleNet (Inception), VGG, ResNet and so on. These networks have been tested by ImageNet, a powerful data set, and are often used in image classification problems.

For the structure of the network, image features are extracted firstly through several layers of CNN, and then nonlinearly classified through the fully connected layer. At this time, the fully connected layer is similar to MLP, but a mechanism such as dropout is also added to prevent overfitting. Etc., there are several classifications in the last layer to connect several neurons, and the probability distribution of the samples belonging to each classification is obtained by softmax transformation.

Discussions on facial expression recognition have continued, and many scholars' teams have focused on this. Document [17] proposes a large database (i.e., a face image downloaded from the Internet) for annotating one million images of natural emotional facial expressions. First, it is proved that this newly proposed algorithm can reliably identify the AU and its strength across databases . According to the survey, this is the first published algorithm to identify high-precision results of AUs and their intensities in multiple databases. The algorithm can run in real time (> 30 images/sec ), allowing it to process large numbers of images and video sequences . Second, use WordNet to download 1,000,000 facial expression images and related emotional keywords from the Internet. These images are then automatically annotated with our AU, AU intensity and sentiment categories by our algorithm . You can get a very useful database that can be used to easily query applications in computer vision, sentiment computing, social and cognitive psychology, and neuroscience using semantic descriptions.

Literature [18] proposes a deep neural architecture that solves these two problems by combining the local and global features of learning in the initial phase, and replicates the message passing algorithm between classes, similar to the graphical model reasoning of the later stages. method. The results show that by increasing the supervision of the end-to-end training model, we have improved the technical level of 5.3% and 8.2% on the BP4D and DISFA data sets respectively based on the existing level.

There is still a lot of discussion based on this research, and interested can be searched.

05 summary

FER's current focus shifts to challenging real-world scenarios, using deep learning techniques to address issues such as lighting changes, occlusion, and non-positive head poses.

Another major issue to consider is that although expression recognition technology has been widely studied, the expressions we define cover only a small part of a particular type, mainly facial expressions, but in fact humans have many other expressions. .

The study of expressions is much more difficult than the age of the face, and the application is much more extensive. I believe that there will be interesting applications in the past few years.
More, welcome to know the column to contribute and communicate.

zhuanlan.zhihu.com/c_15 (Automatic identification of QR code)

github.com/longpeng2008 (Automatic identification of QR code)

Playing a small advertisement, my computer vision open class "AI image recognition project from entry to the line" was launched. It will cover the whole process from zero basis to completing an actual project to the WeChat applet. Welcome to exchange.

References:
[1] He Lianghua. Research on some key technologies in facial expression recognition [D]. Southeast University, 2005.
[2] Zhou Shuren, Liang Ximing, Zhu Can, et al. Expression recognition based on ICA and HMM [J]. Journal of Image and Graphics, 2008, 13(12): 2321-2328.
[3] Zhou Shuren. Analysis and Research of Facial Expression Recognition Algorithm [D]. Central South University, 2009.
[4] Ying Zao, Tang Jinghai, Li Jingwen, Et al. Support vector discrimination analysis and its application in facial expression recognition[J]. Chinese Journal of Electronics, 2008, 36(4): 725-730.
[5] Khorasani K. Facial expression recognition using constructive neural networks[C]// Signal Processing, Sensor Fusion, and Target Recognition X. Signal Processing, Sensor Fusion, and Target Recognition X, 2001:1588 - 1595.
[6] Kyperountas M, Tefas A, Pitas I. Salient feature and reliable classifier selection for facial expression classification [J] PatternRecognition, 2010, 43(3): 972-986.
[7] Zheng W, Zhou X, Zou C, et al. Facial expression recognition using kernel canonical correlation analysis (KCCA). [J]. IEEE Transactions on Neural Networks, 2006, 17(1): 233.
[8] Fu Xiaofeng . Research on Face Recognition and Expression Recognition Based on Binary Mode[D]. Zhejiang University, 2008.
[9] Yacoob Y, Davis L S. Recognizing Human Facial ExpressionsFrom Long Image Sequences Using Optical Flow[J]. PatternAnalysis & Machine Intelligence IEEE Transactions on,
1996, 18 (6): 636-642. [10] Tsalakanidou F, Malassiotis S. Real-time 2D+3D facial action and expression recognition[J]. Pattern Recognition, 2010,43(5):1763- 1775.
[11] Wang J, Yin L. Static topographic modeling for facial expression recognition and analysis [J]. Computer Vision & Image Understanding, 2007, 108(1):19-34.
[12] Sung J, Kim D. Pose-Robust Facial Expression RecognitionUsing View-Based 2D $+$ 3D AAM[J]. IEEE Transactions onSystems, Man, and Cybernetics - Part A: Systems and Humans, 2008, 38(4) :852-866.
[13] Kotsia I, Pitas I. Facial Expression Recognition in ImageSequences Using Geometric Deformation Features and Support Vector Machines [J]. IEEE Transactions on Image Processing APublication of the IEEE Signal Processing Society, 2007, 16(1): 172.
[14] Cohen I, Sebe N, Garg A, et al. Facial expression recognition from video sequences: temporal and static modeling [J]. Computer Vision & Image Understanding, 2003, 91(1–2): 160-187.
[15] Xu Wenhui, Sun Zhengxing. LSVM algorithm for video sequence expression classification[J]. Journal of Computer-Aided Design & Computer Graphics, 2009, 21(4): 000542-553.
[16] Xu Qinzhen, Zhang Pinzheng, Yan Wenjiang, et al. Automatic facial expression classification based on obfuscated cross-support vector machine tree[J]. Journal of Image and Graphics, 2008, 13(7): 1329-1334.
[17] Benitez-Quiroz CF, Srinivasan R, Martinez A M. EmotioNet: An Accurate, Real-Time Algorithm for the Automatic Annotation of a Million Facial Expressions in the Wild [C]// Computer Vision and Pattern Recognition. IEEE, 2016:5562 -5570.
[18] Benitezquiroz CF, Wang Y, Martinez A M. Recognition of Action Units in the Wild with Deep Nets and a New Global-LocalLoss[C]// IEEE International Conference on Computer Vision. IEEE Computer Society, 2017:3990 -3999.

If you want to join us, please leave a message in the background.

More please move

1, my gitchat master class
Peng-Lung of Daren class gitbook.cn
2, AI technical public number, "There are three AI"
This article talk about the history of the computer to play more than a decade of photography mp.weixin.qq.com
3, as well as the photography number, "There are three studios"
[2] On the photography big coffee selfie, selfie mad I only served on the cliff mp.weixin.qq.com